OraCore · Topic ·research

PagedWeight trims MoE memory without tanking quality

PagedWeight dynamically quantizes MoE weights at runtime to trade GPU memory for KV cache headroom.

1 articles in this thread ·Last updated 1d ago·First seen Jul 20, 2026

Timeline

  1. PagedWeight trims MoE memory without tanking quality seed

    PagedWeight dynamically quantizes MoE weights at runtime to trade GPU memory for KV cache headroom.